The bottleneck was never the X-ray
What £20 million of AI at the front of the lung cancer pathway can and cannot buy

The government has committed £20 million to put AI-assisted chest X-ray analysis into every NHS trust in England by 2029. The stated purpose is speed: faster identification of suspected lung cancer, faster diagnosis, faster treatment. It is a reasonable thing to want. It is also, in this specific form, a thing we have now tested — and the most relevant UK trial reported its result twelve weeks ago.

What the £20 million funds

Around half of NHS trusts already use these tools. The funding, drawn from the AI Diagnostic Fund, extends them across the rest of England, with a further £8.1 million to pilot six more AI technologies at thirteen sites. In practice the software acts as a second read of the chest X-ray: it flags findings that may be abnormal, and it can move suspicious films up the radiology reporting queue. Two functions, bundled into one announcement — and the distinction between them turns out to matter a great deal.

The trial that already tested this

In March 2026, the LungIMPACT trial was published in Nature Medicine. Led from University College London Hospitals across five NHS trusts, it examined more than 93,000 chest X-rays requested by primary care and asked a precise question: if AI flags an abnormal film so a radiologist reports it sooner, does the patient reach a lung cancer diagnosis sooner?

The answer was no. AI prioritisation did exactly what it was designed to do at the reporting step — the median time for a radiologist to report a flagged film fell from 47 hours to 34. But the gain stopped there. Among the patients later found to have lung cancer, the median time from chest X-ray to diagnosis was 44 days with AI and 46 days without: a two-day difference that did not reach significance. Referral rates, treatment start times and cancer stage at diagnosis were the same in both arms. The investigators concluded that worklist prioritisation, on its own, should not be expected to accelerate this pathway.

The principal investigator put the reason plainly: the constraint is not the reporting. It is everything that happens after the report — telling the patient, booking the CT, finding the clinic slot, convening the multidisciplinary team meeting.

Detection and prioritisation are not the same claim

It is worth separating the two functions, because they carry different evidence. Detection — a second read that catches a finding a tired human eye may miss — is a real benefit, and LungIMPACT neither tested nor diminished it. Prioritisation — reordering the queue so suspicious films are read first — is what the trial tested, and is where the speed gain was shown not to propagate. The £20 million is being sold mainly on speed. On speed, the evidence is already in, and it is sobering.

Why this lands on the surgical pathway

Here is the part that matters most for those of us at the operating end of this pathway, and it holds whichever way the technology performs.

If detection succeeds, it will find more abnormalities — including cancers that would otherwise have been missed or found later. That is the goal, and it is the right one. But every additional cancer found is a patient who then needs a CT, a clinic appointment, a diagnosis, a multidisciplinary decision and, for a large proportion, an operation. Better detection at the front door does not reduce the work behind the door. It increases it.

And if prioritisation does not shorten the pathway — as the strongest UK trial has now shown — then the delay was never at the front door to begin with. It sat downstream, in the same CT lists, clinics and theatres that more detection will now fill faster.

Both readings arrive in the same place. The front of the pathway is being made quicker and more sensitive. The corridor behind it — diagnostics, clinic capacity, surgical capacity — is not being widened to match.

The question the announcement does not answer

None of this is an argument against AI in radiology. I use image-guided technology in my own practice, and a second read that catches an early cancer is worth having on its own terms. The argument is narrower, and I think harder to dismiss: investment concentrated at the point of detection, without matched investment in the steps that follow detection, does not produce faster treatment. It produces the same queue, entered sooner. A detection gain that downstream capacity cannot absorb becomes a longer waiting list, not a faster diagnosis.

The discipline this asks of us is unglamorous. Before funding a tool that finds disease faster, ask what happens to the patient in the weeks after it is found — and fund that with equal seriousness. The front door of the lung cancer pathway is being rebuilt, and rightly so. It is not the only part being rebuilt. Whether any of it reaches the patient depends on a question this announcement leaves open: who is widening the corridor.

← Back to the Journal